In RL, reward does not mean positive only. You could train an agent on purely negative rewards and it would optimize for the policy that is least negative. Als

In RL, reward does not mean positive only. You could train an agent on purely negative rewards and it would optimize for the policy that is least negative. Also, +1 and -1 do cancel each other out by definition. If your model has suffering greater than succeeding, then you would assign suffering a greater penalty, like -10 vs +1.